TCP and UDP Performance for Internet over Optical Packet-Switched Networks

advertisement
TCP and UDP Performance for Internet over Optical
Packet-Switched Networks
Jingyi He†
S.-H. Gary Chan‡
†
Department of Electrical and Electronic Engineering
‡
Department of Computer Science
Hong Kong University of Science and Technology
Clear Water Bay, Kowloon
Hong Kong
Abstract— Optical packet-switched (OPS) network is a strong candidate for the future Optical Internet. In this paper, we study packet
aggregation and deflection routing as employed in OPS networks on the
performance of upper layer Internet protocols represented by TCP and
UDP. Regarding packet aggregation schemes, we study no aggregation,
mixed-flow aggregation, and per-flow aggregation. Our results show that
for both TCP and UDP, mixed-flow aggregation achieves the highest
throughput, and per-flow aggregation significantly decreases the fairness
at large aggregation intervals. Both aggregation schemes increase UDP
delay jitter. Regarding deflection routing, we show that deflection routing
significantly improves TCP throughput in spite of the out-of-order packet
delivery. However, the congestion of the deflection path significantly
affects the improvement that can be achieved. UDP throughput does
not suffer from out-of-order packet delivery, and hence its throughput
improvement by deflection routing can be even more prominent. The
larger the deflection cost (delay difference between the deflection path
and the shortest path) is, the larger UDP delay jitter results. Deflection
cost, however, does not affect the throughput and fairness very much for
both TCP and UDP.
I. I NTRODUCTION
Driven by the ever-increasing demand for bandwidth, the core
of the Internet has been evolving from an electronic network to
an optical one (i.e., the so-called Optical Internet). Optical packetswitched (OPS) network is a strong candidate for this Optical Internet.
Compared with wavelength-routed network, which is deemed as
another candidate and offers bandwidth granularity only at the wavelength level, OPS network offers bandwidth granularity at the packet
level, resulting in a much higher bandwidth efficiency. Moreover,
because of its packet-switched nature, OPS network can have richer
routing functionalities and greater flexibility in supporting diverse
services [1]. In this paper, we will focus on delivering Internet traffic
over an OPS core network.
In an OPS-based Internet, end users are still attached to electronic
networks. Their traffic is aggregated before being forwarded to
the OPS backbone. Transmission Control Protocol (TCP) and User
Datagram Protocol (UDP) are currently the predominant transport
protocols used at the end users. These packets will eventually be
forwarded in the optical core. In this paper, we explore the impact
of mechanisms as used in the OPS network (namely, routing and
packet handling) on the performance of the Internet protocol suite.
The performance of interest are throughput (goodput), fairness, and,
in case of UDP, delay jitter.
In particular, we investigate the following two mechanisms as
employed in OPS networks which potentially impact the performance
of TCP and UDP:
This work was supported, in part, by the Areas of Excellence (AoE)
Scheme on Information Technology funded by the University Grant Council
in Hong Kong (AoE/E-01/99), and by the Competitive Earmarked Research
Grant of the Hong Kong Research Grant Council (HKUST6014/01E &
HKUST6199/02E).
Packet aggregation — In OPS networks, IP packets destined to
the same destination can be aggregated to form a larger optical
packet. This is necessary especially when the optical switches
are not fast enough to process individual IP packets. Clearly,
packet aggregation implies waiting delay for some packets.
Moreover, an optical packet may contain multiple packets from
the same flow. When it is dropped or mis-routed in the OPS
network, the throughput of the flow may be adversely affected.
In this regard, how packet aggregation affects the TCP and
UDP performance is an important issue. We study in this paper
three aggregation schemes: no aggregation, aggregating packets
from heterogeneous flows into an optical packet (mixed-flow
aggregation), and aggregating packets from the same flow (perflow aggregation).
• Deflection routing — In packet-switched networks, contention
occurs when two or more packets are to be forwarded to the
same output port. In electronic networks, such contention is
resolved in a store-and-forward manner by temporarily buffering
the contending packets in random access memory (RAM).
However, since there is no effective optical RAM, contention
in OPS networks can usually be resolved by deflection routing,
in which the packets losing the contention are temporarily misrouted, or “deflected”, to other output ports.1 Clearly, deflection
routing avoids packet being lost by routing them to longer paths.
Deflection routing also leads to out-of-order delivery of packets.
For a protocol which has reassembly deadline and congestion
control mechanism such as TCP, this will adversely affect its
performance.
In this paper, we study the impact of the aforementioned schemes
as used in unslotted irregular OPS networks on TCP and UDP
performance. Our results show that:
1) For both TCP and UDP, mixed-flow aggregation achieves
the highest throughput, and per-flow aggregation significantly
decreases the fairness at large aggregation intervals. Both
aggregation schemes increases UDP delay jitter.
2) Deflection routing significantly improves TCP throughput in
spite of the out-of-order packet delivery. However, the congestion of the deflection path significantly affects the improvement
that can be achieved. UDP throughput does not suffer from outof-order packet delivery, and hence its throughput improvement
by deflection routing can be more prominent.
3) Deflection cost (delay difference between the deflection path
and the shortest path) does not affect the throughput and
fairness very much for both TCP and UDP, while a larger one
•
1 Another way to resolve contention is wavelength conversion. However,
since it is much more expensive, we would not consider it in this study.
0-7803-7802-4/03/$17.00 (C) 2003 IEEE
Edge OPS
i.e., fixed-interval aggregation. In practice, we may also want to limit
the size of an aggregate packet. In that case, an aggregation interval
expires when the size of the aggregate packet reaches a predetermined
threshold. There can be two ways to aggregate packets:
Edge OPS
E/O
O/E
•
Core OPS
OPS network
IP Subnet
IP Subnet
Fig. 1. An optical packet-switched (OPS) network used to transport traffic
between IP subnetworks.
results in larger UDP delay jitter.
We briefly discuss previous work as follows. Deflection routing has
long been proposed and studied. However, it is mainly investigated
in slotted optical network of regular topology such as the Manhattan
Street Network (MS-Net) and the ShuffleNet [2], [3]. We believe that
in the future Optical Internet, unslotted network with irregular mesh
topology is more likely [4]. There are several pieces of work that
study deflection routing in irregular OPS networks [5], [6]. However,
these studies have not considered the effect of out-of-order packet
delivery by assuming that the receiver has an infinite reassembly time
and buffer. Though the effect of finite reassembly buffer has been
discussed in [7], the study assumes a very simple flow control model
markedly different from TCP, where duplicated acknowledgement,
fast retransmit, and fast recovery are involved. We have studied the
impact of deflection routing on TCP performance in an unslotted
irregular network in [8]. However, the paper has not considered the
special characteristics of OPS networks such as the lack of optical
RAM and the use of packet aggregation. Packet aggregation was
previously studied mainly in the context of web traffic in the Internet
[9], [10].
The rest of this paper is organized as follows. We present in
Section II the system description of the OPS network under study
and our simulation model. In Section III, we present some illustrative
simulation results of the impact of packet aggregation on TCP and
UDP performance. In Section IV, we present the effect of deflection
routing. We conclude in Section V.
II. S YSTEM D ESCRIPTION AND S IMULATION M ODEL
A. System Description
We consider that an OPS network is used to transport the traffic
between IP subnetworks, as shown in Fig. 1. The OPS network
consists of edge OPSes and core OPSes inter-connected by optical
links. The core OPSes are responsible for route selection and label
swapping for the optical packets. The edge OPSes interface with IP
subnetworks, and are responsible for packet classification, aggregation, and segregation, in addition to the switching functionalities.
An IP packet encounters two edge OPSes on its path to the
destination: an ingress OPS where it enters the OPS network and an
egress OPS where it leaves the network. At the ingress OPS, the IP
packet may firstly be buffered and then aggregated with other packets
to form a larger packet. These aggregated IP packets are transmitted
into the OPS network after an addition of an optical packet header and
electronic-optical (E/O) conversion. When the optical packet reaches
the egress OPS, it is segregated through optical-electronic (O/E)
conversion into individual IP packets, which are then transmitted to
their respective destinations.
In terms of packet aggregation, we consider a scheme in which an
aggregate packet is formed when a fixed aggregation interval expires,
•
Mixed-flow aggregation: In this scheme, IP packets from different flows may be aggregated into an optical packet as long
as they are destined to the same egress OPS. Therefore, at an
ingress OPS, only one aggregation timer is needed for each
egress OPS.
Per-flow aggregation: In this scheme, an optical packet can
only consist of IP packets from the same flow. For each active
flow, an aggregation timer is needed, which starts from the first
packet of the flow at the ingress OPS. This scheme makes
the manipulation of individual flows possible. However, it is
clearly more complex: IP packets are to be further differentiated
according to their flow identification and per-flow queueing is
needed at the ingress OPS.
We next discuss the condition under which deflection occurs. Note
that a newly-formed optical packet at the ingress OPS may not be
able to be sent out to its preferred outgoing port right away, because
the port is occupied by another optical packet in transmission. For
example, if the bandwidth of the outgoing link is B, the size of an
earlier optical packet is S, and the interval between the formation
time of the new optical packet and that of the earlier packet is τ , the
new packet will not be able to be forwarded to the same outgoing port
if τ < S/B. In this case, deflection routing can be used to resolve
the contention. At the core OPSes, without optical buffer, deflection
happens when τ < S/B, where τ is the inter-arrival time between
two optical packets contending for the same output link.
B. Simulation Model
The simulation model we use is shown in Fig. 2. We focus on a set
of flows which are to be transported by an OPS network, and enter
at the same ingress switch r0 and leave at the same egress switch
r1 . The flows are sourced at nodes si (i = 1, 2, . . . , n) and destined
to nodes di (i = 1, 2, . . . , n). We consider that for each si − di pair
there are m flows. The link connecting si with r0 and that connecting
di with r1 have bandwidth Bi and delay ti . All the links inside the
OPS network have bandwidth B0 . The delay of the shortest path
between r0 and r1 is assumed to be t0 (a constant). The optical
packets deflected away from the shortest path will be subjected to
a longer (random) delay of t0 + tc , where tc is a positive random
variable indicating the deflection cost. At the ingress OPS r0 , the IP
packets are aggregated with an aggregation interval ∆, and size limit
B0 ∆.
After an aggregate packet enters the OPS network, it may be
deflected at different points and for different number of times,
and some of them may even be dropped because of the ongoing
background traffic on their deflection paths. In order to capture this
situation properly for a general network, we introduce in our model
a parameter p, which is defined as the probability that an optical
packet deflected at its ingress OPS is dropped on the deflection path.
Clearly, p depends on the traffic load on the deflection path: p is
small if there is little background traffic and large if there is heavy
traffic on the deflection path. Hence we call p the congestion factor
of the deflection path. For simplicity, we assume that the packets
forwarded to the shortest path are not dropped.
In the reverse direction, r1 acts as the ingress OPS and r0 the
egress OPS. The traffic in this direction (e.g., TCP ACKs) is treated
in the same manner as the traffic in the forward direction.
0-7803-7802-4/03/$17.00 (C) 2003 IEEE
Notation: (bandwidth, delay)
600
r2
0.95
550
0.9
No aggregation
500
d1
450
s2
Throughput (Mbps)
( B1, t 1)
( B1, t 1)
d2
Deflection paths ( B0, t 0 + t c )
.
.
.
r0
Shortest path
( B0, t 0)
400
350
OPS Network
( Bn, t n)
No aggregation
0
10
Aggregation interval ∆ (ms)
(a) Throughput
( Bn, t n)
dn
Fig. 2.
0.65
200
dn-1
sn
Per−flow aggregation
250
150 −1
10
sn-1
0.8
0.75
0.7
300
.
.
.
r1
Mixed−flow aggregation
Fairness index
s1
0.85
Mixed−flow aggregation
Fig. 3.
Per−flow aggregation
0.6
1
10
0.55 −1
10
0
10
1
10
Aggregation interval ∆ (ms)
(b) Fairness index
TCP performance versus ∆ with different aggregation schemes.
The simulation model we use.
flow. We define the overall delay jitter as the average jitter over all
flows.
TABLE I
BASELINE PARAMETERS USED IN THE SIMULATION .
III. E FFECT OF PACKET AGGREGATION
Symbol
n
m
p
B0
Bi
L
t0
ti
tv
∆
µ
σ
S0
Definition
Number of source and destination nodes
Number of flows for each si − di pair
Congestion factor of the deflection paths
Bandwidth of the links in the OPS network
Bandwidth of the links outside the OPS network
Offered UDP load
Delay of the shortest path between r0 and r1
Delay of the links outside the OPS network
Interval of the delay changes
Aggregation interval
Mean of the deflection cost
Standard deviation of the deflection cost
IP Packet size
Default value
10
50
0.2
622 Mbps
155 Mbps
800 Mbps
10 ms
2i ms
[1, 100] ms
1 ms
5 ms
4 ms
1500 bytes
In this section, we present the illustrative results of TCP and UDP
performance over the OPS network, with different packet aggregation
schemes employed: no aggregation, mixed-flow aggregation, and perflow aggregation. As aggregation is performed every fixed interval,
we study the impact of using different interval values.
A. TCP Performance
We use the ns-2 network simulator to do our simulation [11].2 In
the simulation, the aggregation interval ∆ varies from 0.1 ms to 10
ms, as most of the up-to-date all-optical switches have a switching
speed in the order of milliseconds. We let tc be a truncated Gaussian
random variable with mean µ and standard deviation σ, i.e., tc =
max (0, N (µ, σ)). Since the background traffic on the deflection
paths may vary with time, we let the delay of the deflection path
t0 + tc change every tv time, where tv is uniformly distributed over
a certain time interval, for example, [1, 100] ms. In the simulation, all
the flows have the same packet size of S0 = 1500 bytes. The starting
times of the flows are uniformly distributed in a time interval (e.g.,
[0.1, 0.5] s) to avoid global synchronization. For TCP flows, we use
TCP-Reno and bulk data transfer (e.g., FTP) traffic. For UDP, the
packets of each flow arrive according to a Poisson process, with an
overall traffic load of L. The default (baseline) parameters used in
our simulation are summarized in Table I.
We are mainly interested in the throughput, fairness among the
flows, and packet delay jitters (of UDP flows). In terms of fairness,
we use the Jain’s fairness index defined as [12]:
f=
N
N
x
2
i=1 i
N
x2
i=1 i
,
(1)
where N is the total number of flows and xi is the throughput of
flow i, i = 1, . . . , N . The value of Jain’s fairness index is always
between zero and one, i.e., f ∈ (0, 1]. A larger value means better
fairness, with f = 1 meaning perfect fairness. In terms of delay jitter
of a UDP flow, we use the standard deviation of packet delay of that
2 Note that ns-2 is for simulating store-and-forward networks, while in
an OPS network without optical buffer, optical packets are transmitted in
a cut-through manner without any queueing delay at intermediate nodes.
Modifications are hence needed to account for this characteristic.
We show in Fig. 3 the TCP performance with respect to the
aggregation interval ∆. The case without aggregation (a flat line
in both of the figures) is also shown for comparison. We first look
at the overall TCP throughput as shown in Fig. 3(a). Mixed-flow
aggregation achieves much higher TCP throughput than per-flow
aggregation and no aggregation. Per-flow aggregation achieves higher
throughput than no aggregation only at large aggregation intervals.
For both mixed-flow aggregation and per-flow aggregation, TCP
throughput increases with ∆.
TCP throughput is affected by packet aggregation in the following
ways. First, mixed-flow aggregation reduces the contention at the
ingress OPS by regulating the incoming traffic, resulting in higher
throughput. Without packet aggregation, the contention at the ingress
OPS is significant because of the bursty nature of TCP traffic, and remarkable packet losses result. Mixed-flow aggregation aggregates the
packets and forwards them to the optical switch every fixed interval.
Therefore, the traffic to be switched becomes much more regular,
and packet losses are significantly reduced. Per-flow aggregation,
however, does not help reduce the contention, because the aggregation
timers of the flows are not synchronized, and hence aggregated
packets from different flows are forwarded to the optical switch
still in a random manner. Even worse is that with larger (aggregate)
packet size an idle period is likely larger and hence the bandwidth
is less efficiently used. We will show how this effect decreases the
throughput in UDP case later. For the TCP case here, however, this
effect is not very severe, because the traffic load is relatively low (as
indicated by the throughput) and hence the aggregate packet size is
not large. Actually, we observe that per-flow aggregation achieves
even higher throughput than no aggregation at large aggregation
intervals. The major reasons are less TCP retransmissions and unfair
share of bandwidth among the flows, as will be explained in the
following.
Packet aggregation reduces TCP retransmissions when deflection
routing is used, hence improving TCP throughput. With deflection
routing, a deflected TCP packet will arrive late at its destination
and possibly trigger TCP fast retransmit (if more than three packets
of larger sequence numbers have arrived through the shortest path).
Given the burstiness of TCP traffic, a mixed-flow aggregate packet
0-7803-7802-4/03/$17.00 (C) 2003 IEEE
740
B. UDP Performance
We show in Fig. 4 UDP throughput and fairness versus ∆, with
an overall traffic load L = 800 Mbps. Regarding throughput, like the
TCP case, mixed-flow aggregation achieves higher values than perflow aggregation and no aggregation, because it reduces contention
and hence packet losses. On the other hand, unlike TCP, per-flow
aggregation achieves lower throughput than no aggregation. The
major reason is the inefficient use of bandwidth because of the large
aggregate packet size (given the high traffic load here), as we have
discussed earlier. We also observe that UDP throughput does not show
much dependence on ∆. The reason is that UDP does not go through
back-off mechanism as in TCP for out-of-order packet delivery, hence
the effect of larger ∆ in reducing retransmissions for TCP does not
apply here.
Regarding fairness, both mixed-flow aggregation and no aggregation achieve almost perfect fairness. Per-flow aggregation, however,
achieves fairness degrading as ∆ increases. This is because all the
flows share the same ∆, and hence a flow once suffers from a
deflection loss of its optical packets would more likely suffer from
such losses again later. As a result, some flows suffer from packet
loss more often than others, leading to an unfair share of bandwidth.
As ∆ increases, the fairness decreases because the loss from one
interval to another becomes more correlated.
Packet delay jitter is another important performance metric for
UDP traffic. We show in Fig. 5 the average delay jitter over all
the UDP flows versus ∆. Packet aggregation introduces aggregation
delay which increases the delay jitter. The larger ∆ is, the larger delay
1
Mixed−flow aggregation
0.99
720
Mixed−flow aggregation
No aggregation
0.98
700
0.97
680
0.96
Fairness index
Throughput (Mbps)
660
640
0.95
0.94
Per−flow aggregation
620
0.93
600
580
0.92
No aggregation
0.91
Per−flow aggregation
560 −1
10
0
0.9 −1
10
1
10
10
0
10
Aggregation interval ∆ (ms)
Fig. 4.
1
10
Aggregation interval ∆ (ms)
(b) Fairness index
(a) Throughput
UDP performance versus ∆ with different aggregation schemes.
5
4.8
Average delay jitter (ms)
is likely to consist of packets from only a certain number of flows,
with multiple packets for each flow. As the multiple packets from
the same flow are in order among themselves, when an aggregate
packet is deflected, the block of in-order packets which will arrive
late in general does not cause more TCP retransmissions than a single
deflected TCP packet. This is also true for per-flow aggregation.
With a larger ∆, deflection occurs less frequently, resulting in fewer
retransmissions. Therefore, the TCP congestion window can be kept
at a larger value on average, as well as there are fewer duplicated
packets wasting the bandwidth. The overall effect is larger TCP
throughput for larger ∆. Of course, ∆ could not be too large either,
because that would result in too large delay for some packets, which
may break the TCP pipelining transmission.
We next look at the TCP fairness as shown in Fig. 3(b). In general, packet aggregation degrades the TCP fairness. This is because
packet aggregation does not pack packets evenly from all the flows,
given the burstiness of TCP traffic. With (aggregate) packet losses
and deflections, the unfairness among the flows may therefore be
worsened. With mixed-flow aggregation, TCP fairness index slightly
increases with ∆. This is reasonable because with a larger ∆ an
aggregate packet consists of packets from more flows. When it is
deflected or dropped, the effect is more evenly distributed among
all the flows. For per-flow aggregation, however, TCP fairness index
decreases when ∆ is large. This is because large ∆ leads to large
aggregate packets. As the TCP packets in an aggregate packet are
all from the same flow, a loss of such an aggregate packet easily
leads to TCP timeout. Therefore, some flows are likely severely
starved while giving bandwidth to others, resulting in even worse
fairness. In such cases, however, the overall TCP throughput can be
even larger, because a small number of flows occupying most of
the bandwidth means that these flows suffer few packet losses and
retransmissions, and hence most of the bandwidth accounts for the
throughput (goodput). This has been demonstrated in Fig. 3(a).
Per−flow aggregation
4.6
4.4
Mixed−flow aggregation
4.2
4
No aggregation
3.8 −1
10
0
10
1
10
Aggregation interval ∆ (ms)
Fig. 5. UDP packet delay jitter versus ∆ with different aggregation schemes.
jitter results. With the given load L = 800 Mbps, the packet arrival
rate of each flow is one packet per 7.5 ms. For per-flow aggregation,
when the aggregation interval ∆ is small, in most cases an aggregate
packet will consists of only one IP packet. In this case, nearly all
the deflected packets are subjected to the same aggregation delay
(statistically). Therefore, the delay jitter resulting from aggregation
delay is small. When ∆ is larger, the probability that an aggregate
packet consists of multiple IP packets is larger. In such case, the
IP packets in an aggregate packet suffer much different aggregation
delays, resulting a fast increase in the overall delay jitter. For mixedflow aggregation, with the given traffic load, an aggregate packet is
generally formed and forwarded when its size threshold is reached
before the aggregation interval expires. In this case, the randomness
of aggregation delay is larger than that of per-flow aggregation when
∆ is small. We hence observe that mixed-flow aggregation results
in larger delay jitter than per-flow aggregation when ∆ is not very
large. On the other hand, when ∆ is large, this mechanism results
in significant reduction in aggregation delay compared with the fixed
aggregation interval. Therefore, mixed-flow aggregation can achieve
a smaller delay jitter than per-flow aggregation when ∆ is large.
IV. E FFECT OF D EFLECTION ROUTING
In this section, we study the effect of deflection routing on TCP and
UDP performance. Specifically, we study the influence of deflection
cost (mean µ and standard deviation σ) and the congestion factor p,
with mixed-flow aggregation and ∆ = 1 ms.
We show in Fig. 6 TCP throughput and fairness index versus the
mean deflection cost µ. The case without deflection (the flat line)
is also shown for comparison. We can see that deflection routing
significantly improves TCP throughput (∼ 20%), although it results
in out-of-order packet delivery. The reason is that packet loss is more
harmful to TCP than out-of-order packet delivery. Moreover, deflection routing also improves TCP fairness. As can be seen, however, the
throughput and fairness index do not show much dependency on µ. In
an OPS network without optical buffer, there is no queueing delay at
intermediate nodes, and hence a deflected aggregate packet will arrive
(at the egress OPS) later than its following one along the shortest path
as long as its deflection cost is larger than the aggregation interval.
0-7803-7802-4/03/$17.00 (C) 2003 IEEE
520
0.88
800
1
0.87
750
0.98
UDP
500
Deflection
0.86
0.96
700
Deflection
0.85
0.84
0.83
650
Fairness index
460
Throughput (Mbps)
0.94
Fairness index
Throughput (Mbps)
480
UDP
600
550
440
0.82
500
No deflection
No deflection
0.86
TCP
2.5
5
7.5
10
12.5
0.8
0
15
2.5
5
Mean of deflection cost µ (ms)
12.5
15
9
9
8
8
7
7
6
5
5
4
3
3
5
7.5
10
12.5
Mean of deflection cost µ (ms)
2
0
2
4
6
8
10
12
Standard deviation of deflection cost σ (ms)
(a) Delay jitter vs. µ
Fig. 7.
15
0.84
0.1
0.2
0.3
0.4
0.5
0.6
0.7
(a) Throughput
6
4
400
0
0.8
0.9
1
0.82
0
0.1
Congestion factor of deflection path p
(b) Fairness index
Average delay jitter (ms)
Average delay jitter (ms)
10
TCP performance versus the mean of deflection cost µ.
Fig. 6.
2.5
7.5
Mean of deflection cost µ (ms)
(a) Throughput
TCP
450
0.81
2
0
0.9
0.88
420
400
0
0.92
(b) Delay jitter vs. σ
UDP delay jitter versus µ and σ.
This is in general true in our study, given ∆ = 1 ms and σ = 4 ms.
Therefore, in terms of triggering TCP fast retransmit, larger µ does
not make much difference. We observe a slightly decreasing trend
of throughput, which is mainly because larger µ results in larger
average round-trip time (RTT). Our simulation results also show that
the standard deviation of the deflection cost σ does not affect TCP
throughput and fairness either.
With or without deflection, UDP can always achieve very good
fairness. Regarding throughput, as UDP does not have any backoff mechanism in response to out-of-order packet delivery, deflected
packets directly account for throughput improvement, and the improvement can be even more prominent than TCP. For the same
reason, UDP throughput is independent of µ and σ. However, UDP
delay jitter performance is affected by out-of-order packet delivery.
Our simulation results indicate that the average delay jitter increases
linearly with both µ and σ, as shown in Fig. 7.
We finally show in Fig. 8 the effect of p on TCP and UDP
performance. As expected, larger p results in more packet losses
and hence lower throughput for both TCP and UDP. For UDP, the
throughput decreases linearly with respect to p, because the amount
of packet losses is proportional to p. For TCP, in contrast, the
throughput drops sharply as p increases from 0 to 0.1, and then keeps
on decreasing more mildly. This actually means that packet loss is
very harmful to TCP throughput, and the loss of deflected packet
significantly decreases the throughput improvement by deflection
routing. Regarding fairness, UDP can achieve almost perfect fairness
irrespective of p. For TCP, the fairness index shows an obvious
drop when p increases from 0 to 0.1, similar to the throughput
performance. The increase of p beyond that does not show much
effect on the fairness index. Our results also show that larger p
leads to smaller UDP delay jitter, because fewer deflected packets
are received.
V. C ONCLUSIONS
Optical packet-switched (OPS) networks will likely carry Internet
traffic in the future. In this paper, we have examined the performance
of Internet protocols (TCP and UDP) over OPS networks. In particular, we have studied the impact of packet aggregation and deflection
Fig. 8.
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Congestion factor of deflection path p
(b) Fairness index
TCP and UDP performance versus p.
routing as used in the underlying OPS networks on TCP/UDP
performance. The performance metrics we are interested in are
throughput, fairness, and delay jitter. Regarding packet aggregation
schemes, we have studied no aggregation, mixed-flow aggregation
and per-flow aggregation. Our results show that for both TCP and
UDP, mixed-flow aggregation achieves the highest throughput, and
per-flow aggregation significantly decreases the fairness at large
aggregation intervals. Both aggregation schemes increases UDP delay
jitter. Regarding deflection routing, we show that deflection routing
significantly improves TCP throughput in spite of the out-of-order
packet delivery. However, the congestion of the deflection path
significantly affects the improvement that can be achieved. UDP
throughput does not suffer from out-of-order packet delivery, and
hence its throughput improvement by deflection routing can be even
more prominent. Deflection cost does not affect the throughput and
fairness very much for both TCP and UDP, while larger deflection
cost results in larger UDP delay jitter.
R EFERENCES
[1] M. J. O’Mahony, D. Simeonidou, D. K. Hunter, and A. Tzanakaki,
“The application of optical packet switching in future communication
networks,” IEEE Communications Magazine, vol. 39, no. 3, pp. 128–
135, Mar. 2001.
[2] A. K. Choudhury, “A comparative study of architectures for deflection
routing,” in Proc. IEEE GLOBECOM’92, pp. 1911–1920, Dec. 1992.
[3] S.-H. G. Chan and H. Kobayashi, “Packet scheduling algorithms and
performance of a buffered shufflenet with deflection routing,” IEEE/OSA
Journal of Lightwave Technology, vol. 18, no. 4, pp. 490–501, Apr. 2000.
[4] T. Chich, J. Cohen, and P. Fraigniaud, “Unslotted deflection routing: a practical and efficient protocol for multihop optical networks,”
IEEE/ACM Transactions on Networking, vol. 9, no. 1, pp. 47–59, Feb.
2001.
[5] S. Yao, B. Mukherjee, S. J. B. Yoo, and S. Dixit, “All-optical packetswitched networks: A study of contention-resolution schemes in an
irregular mesh network with variable-sized packets,” in Proc. OptiComm
2000, pp. 235–246, Nov. 2000.
[6] J. P. Jue, “An algorithm for loopless deflection in photonic packetswitched networks,” in Proc. IEEE ICC’2002, pp. 2776 –2780, Apr.
2002.
[7] A. K. Choudhury and N. F. Maxemchuk, “Effect of a finite reassembly
buffer on the performance of deflection routing,” in Proc. IEEE ICC’91,
pp. 1637–1646, Jun. 1991.
[8] J. He and S.-H. G. Chan, “TCP performance with deflection routing in
the Internet,” in Proc. IEEE ICON’2002, pp. 383–388, Aug. 2002.
[9] B. R. Badrinath and P. Sudame, “Gathercast: The design and implementation of a programmable aggregation mechanism for the Internet,”
in Proc. Ninth International Conference on Computer Communications
and Networks, pp. 206–213, 2000.
[10] H. Tounsi, L. Toutain, and F. Kamoun, “Small packets aggregation in
an IP domain,” in Proc. Sixth IEEE Symposium on Computers and
Communications, pp. 708–713, 2001.
[11] “http://www.isi.edu/nsnam/ns.”
[12] R. Jain, D. Chiu, and W. Hawe, “A quantitative measure of fairness and
discrimination for resource allocation in shared computer systems,” in
DEC Research Report TR-301, Sept. 1984.
0-7803-7802-4/03/$17.00 (C) 2003 IEEE
Download